Aprendizado de máquina aplicado à predição de viabilidade técnica para conexão de cargas em rede de distribuição de média tensão

Abstract

The electrical system is being demanded by a growing volume of connection requests by users of distributed generation, industrial, commercial and residential, which makes the analysis process complex, reflecting in an extension in the response time. Carrying out a network mapping with an estimate of maximum power for the service makes the user's decision process simpler and more efficient, benefiting the user and the distributor. The proposed work presents a study focused on predicting the power available for connection of loads at any point of the medium voltage electrical network in the state of Ceará without the need for interventions by the distributor. Predictive analysis was performed using Random Forest (RF), Support Vector Machine (SVM), Artificial Neural Networks (ANN) and Extreme Gradient Boosting (XGBoost) algorithms. The machine learning models were applied in an electrical system database and in power flow studies by a specialist from Enel Distribuição Ceará. The definition of the best algorithm was based on performance metrics, resulting in XGBoost with an overall hit rate of 90%. The results demonstrated the efficiency of the method in estimating power with the predictors available in the four models obtained. An interactive map prototype was created with XGBoost in order to facilitate the interface between the user and predictions, with two versions, one presenting a view of all the feeder bars and another simpler with only the substations.